NEULP: An End-to-End Deep-Learning Model for Link Prediction

Graph neural networks (GNNs) have been proved useful for link prediction in online social networks. However, existing GNNs can only adopt shallow architectures as too many layers will lead to over-smoothing and vanishing gradient during training. It causes nodes to have indistinguishable embeddings if locally they have similar structural positions, and this further leads to inaccurate link prediction. In this paper, we propose a unified end-to-end deep learning model, namely Neural Link Prediction (NEULP), which can integrate the linearity and non-linearity user interactions to overcome the limitation of GNNs. The experimental evaluation demonstrates our model's significant improvement over several baseline models. Moreover, NeuLP achieves a reliable link prediction given two users' different types of attributes and it can be applied to other pairwise tasks. We further perform in-depth analyses on the relation between prediction performance and users' geodesic distance and show that NeuLP still can make accurate link prediction while two users are far apart in the networks.